基于迁移深度学习的雷达信号分选识别

Radar Signal Sorting and Recognition Based on Transferred Deep Learning

针对当前雷达信号分选识别算法普遍存在的低信噪比下识别能力差、特征参数提取困难、分类器模型参数复杂等问题,提出了一种基于时频分析、深度学习和迁移学习融合模型的雷达信号自动分选识别算法。首先通过引入的多重同步压缩变换得到雷达信号的时频图像,然后利用灰度化、维纳滤波、双三次插值法和归一化等手段对时频图像进行预处理,最后基于迁移学习的方法,以GoogLeNet和ResNet模型为基础完成了对雷达信号的离线训练和在线识别。仿真结果表明,在信噪比为?6 dB时,该算法对9种雷达信号(CW, LFM, NLFM, BPSK, MPSK, Costas, LFM/BPSK, LFM/FSK, BPSK/FSK)的整体平均识别率可达93.4%,较常规人工提取算法具有更好的抗噪性和泛化能力。

Aiming at the problems of poor recognition ability under low signal-to-noise ratio(SNR),difficulty in extracting feature parameters and complexity of classifier model parameters commonly existing in current radar signal sorting and recognition algorithms,an automatic radar signal sorting and recognition algorithm based on time-frequency analysis,deep learning and transfer learning fusion model is proposed.Firstly,the time-frequency image of radar signal is obtained by introducing Mul-tisynchrosqueezing Transform.Then,the time-frequency image is preprocessed by gray scale,Wiener filtering,bicubic interpolation and normalization.Finally,based on the migration learning method,the off-line training and on-line recognition of radar signals are completed on the basis of Goog-LeNet and ResNet models.Simulation results show that when SNR is?6 dB,the overall average recognition rate of the algorithm for nine radar signals(CW,LFM,NLFM,BPSK,MPSK,Costas,LFM/BPSK,LFM/FSK,BPSK/FSK)can reach 93.4%,which is better than the conventionally artificial extraction algorithm in noise resistance and generalization.